Uni-Cheb: A Basis-Agnostic Learnable Chebyshev Filter for Multimodal Spectral Modulation

While spatiotemporal modeling has emerged as an effective approach in multimodal learning, it struggles to resolve the ``spectral dilemma'' where distinct tasks demand contradictory frequency views. For example, deception detection requires global macro-patterns, while intent recognition depends on local micro-transients. To address this, we propose \textbf{Uni-Cheb}, a unified spectral operator centered on the Learnable Chebyshev Filter (LCF). Designed as a basis-agnostic, plug-and-play component, LCF maintains a consistent mathematical form to adaptively modulate frequency components regardless of the underlying spectral transform. By leveraging the minimax property of Chebyshev approximation, LCF performs elastic spectral resampling to focus on task-relevant bands. Guided by task-specific physical priors, the LCF seamlessly integrates with the Discrete Fourier Transform (DFT) to amplify global physiological rhythms or the Discrete Wavelet Transform (DWT) to capture localized semantic shifts, all while maintaining a negligible computational footprint. Furthermore, it serves as an effective spectral preconditioner to mitigate distribution shifts in cross-domain transfer learning. Extensive experiments across multiple benchmarks demonstrate that Uni-Cheb acts as a universal enhancer, consistently improving state-of-the-art baselines. By rendering frequency modulation both task-adaptive and highly efficient, Uni-Cheb establishes a robust, operator-level solution to the spectral dilemma.